From 495ffc4777522e40941753e3b1b79c02f84b25b4 Mon Sep 17 00:00:00 2001 From: Grafting Rayman <156515434+GraftingRayman@users.noreply.github.com> Date: Fri, 17 Jan 2025 11:00:30 +0000 Subject: Add files via upload --- r_basicsr/ops/dcn/__init__.py | 7 + r_basicsr/ops/dcn/deform_conv.py | 379 ++++++++++ r_basicsr/ops/dcn/src/deform_conv_cuda.cpp | 685 ++++++++++++++++++ r_basicsr/ops/dcn/src/deform_conv_cuda_kernel.cu | 867 +++++++++++++++++++++++ r_basicsr/ops/dcn/src/deform_conv_ext.cpp | 164 +++++ 5 files changed, 2102 insertions(+) create mode 100644 r_basicsr/ops/dcn/__init__.py create mode 100644 r_basicsr/ops/dcn/deform_conv.py create mode 100644 r_basicsr/ops/dcn/src/deform_conv_cuda.cpp create mode 100644 r_basicsr/ops/dcn/src/deform_conv_cuda_kernel.cu create mode 100644 r_basicsr/ops/dcn/src/deform_conv_ext.cpp (limited to 'r_basicsr/ops/dcn') diff --git a/r_basicsr/ops/dcn/__init__.py b/r_basicsr/ops/dcn/__init__.py new file mode 100644 index 0000000..b534fc6 --- /dev/null +++ b/r_basicsr/ops/dcn/__init__.py @@ -0,0 +1,7 @@ +from .deform_conv import (DeformConv, DeformConvPack, ModulatedDeformConv, ModulatedDeformConvPack, deform_conv, + modulated_deform_conv) + +__all__ = [ + 'DeformConv', 'DeformConvPack', 'ModulatedDeformConv', 'ModulatedDeformConvPack', 'deform_conv', + 'modulated_deform_conv' +] diff --git a/r_basicsr/ops/dcn/deform_conv.py b/r_basicsr/ops/dcn/deform_conv.py new file mode 100644 index 0000000..32de9ef --- /dev/null +++ b/r_basicsr/ops/dcn/deform_conv.py @@ -0,0 +1,379 @@ +import math +import os +import torch +from torch import nn as nn +from torch.autograd import Function +from torch.autograd.function import once_differentiable +from torch.nn import functional as F +from torch.nn.modules.utils import _pair, _single + +BASICSR_JIT = os.getenv('BASICSR_JIT') +if BASICSR_JIT == 'True': + from torch.utils.cpp_extension import load + module_path = os.path.dirname(__file__) + deform_conv_ext = load( + 'deform_conv', + sources=[ + os.path.join(module_path, 'src', 'deform_conv_ext.cpp'), + os.path.join(module_path, 'src', 'deform_conv_cuda.cpp'), + os.path.join(module_path, 'src', 'deform_conv_cuda_kernel.cu'), + ], + ) +else: + try: + from . import deform_conv_ext + except ImportError: + pass + # avoid annoying print output + # print(f'Cannot import deform_conv_ext. Error: {error}. You may need to: \n ' + # '1. compile with BASICSR_EXT=True. or\n ' + # '2. set BASICSR_JIT=True during running') + + +class DeformConvFunction(Function): + + @staticmethod + def forward(ctx, + input, + offset, + weight, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + im2col_step=64): + if input is not None and input.dim() != 4: + raise ValueError(f'Expected 4D tensor as input, got {input.dim()}D tensor instead.') + ctx.stride = _pair(stride) + ctx.padding = _pair(padding) + ctx.dilation = _pair(dilation) + ctx.groups = groups + ctx.deformable_groups = deformable_groups + ctx.im2col_step = im2col_step + + ctx.save_for_backward(input, offset, weight) + + output = input.new_empty(DeformConvFunction._output_size(input, weight, ctx.padding, ctx.dilation, ctx.stride)) + + ctx.bufs_ = [input.new_empty(0), input.new_empty(0)] # columns, ones + + if not input.is_cuda: + raise NotImplementedError + else: + cur_im2col_step = min(ctx.im2col_step, input.shape[0]) + assert (input.shape[0] % cur_im2col_step) == 0, 'im2col step must divide batchsize' + deform_conv_ext.deform_conv_forward(input, weight, + offset, output, ctx.bufs_[0], ctx.bufs_[1], weight.size(3), + weight.size(2), ctx.stride[1], ctx.stride[0], ctx.padding[1], + ctx.padding[0], ctx.dilation[1], ctx.dilation[0], ctx.groups, + ctx.deformable_groups, cur_im2col_step) + return output + + @staticmethod + @once_differentiable + def backward(ctx, grad_output): + input, offset, weight = ctx.saved_tensors + + grad_input = grad_offset = grad_weight = None + + if not grad_output.is_cuda: + raise NotImplementedError + else: + cur_im2col_step = min(ctx.im2col_step, input.shape[0]) + assert (input.shape[0] % cur_im2col_step) == 0, 'im2col step must divide batchsize' + + if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]: + grad_input = torch.zeros_like(input) + grad_offset = torch.zeros_like(offset) + deform_conv_ext.deform_conv_backward_input(input, offset, grad_output, grad_input, + grad_offset, weight, ctx.bufs_[0], weight.size(3), + weight.size(2), ctx.stride[1], ctx.stride[0], ctx.padding[1], + ctx.padding[0], ctx.dilation[1], ctx.dilation[0], ctx.groups, + ctx.deformable_groups, cur_im2col_step) + + if ctx.needs_input_grad[2]: + grad_weight = torch.zeros_like(weight) + deform_conv_ext.deform_conv_backward_parameters(input, offset, grad_output, grad_weight, + ctx.bufs_[0], ctx.bufs_[1], weight.size(3), + weight.size(2), ctx.stride[1], ctx.stride[0], + ctx.padding[1], ctx.padding[0], ctx.dilation[1], + ctx.dilation[0], ctx.groups, ctx.deformable_groups, 1, + cur_im2col_step) + + return (grad_input, grad_offset, grad_weight, None, None, None, None, None) + + @staticmethod + def _output_size(input, weight, padding, dilation, stride): + channels = weight.size(0) + output_size = (input.size(0), channels) + for d in range(input.dim() - 2): + in_size = input.size(d + 2) + pad = padding[d] + kernel = dilation[d] * (weight.size(d + 2) - 1) + 1 + stride_ = stride[d] + output_size += ((in_size + (2 * pad) - kernel) // stride_ + 1, ) + if not all(map(lambda s: s > 0, output_size)): + raise ValueError(f'convolution input is too small (output would be {"x".join(map(str, output_size))})') + return output_size + + +class ModulatedDeformConvFunction(Function): + + @staticmethod + def forward(ctx, + input, + offset, + mask, + weight, + bias=None, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1): + ctx.stride = stride + ctx.padding = padding + ctx.dilation = dilation + ctx.groups = groups + ctx.deformable_groups = deformable_groups + ctx.with_bias = bias is not None + if not ctx.with_bias: + bias = input.new_empty(1) # fake tensor + if not input.is_cuda: + raise NotImplementedError + if weight.requires_grad or mask.requires_grad or offset.requires_grad or input.requires_grad: + ctx.save_for_backward(input, offset, mask, weight, bias) + output = input.new_empty(ModulatedDeformConvFunction._infer_shape(ctx, input, weight)) + ctx._bufs = [input.new_empty(0), input.new_empty(0)] + deform_conv_ext.modulated_deform_conv_forward(input, weight, bias, ctx._bufs[0], offset, mask, output, + ctx._bufs[1], weight.shape[2], weight.shape[3], ctx.stride, + ctx.stride, ctx.padding, ctx.padding, ctx.dilation, ctx.dilation, + ctx.groups, ctx.deformable_groups, ctx.with_bias) + return output + + @staticmethod + @once_differentiable + def backward(ctx, grad_output): + if not grad_output.is_cuda: + raise NotImplementedError + input, offset, mask, weight, bias = ctx.saved_tensors + grad_input = torch.zeros_like(input) + grad_offset = torch.zeros_like(offset) + grad_mask = torch.zeros_like(mask) + grad_weight = torch.zeros_like(weight) + grad_bias = torch.zeros_like(bias) + deform_conv_ext.modulated_deform_conv_backward(input, weight, bias, ctx._bufs[0], offset, mask, ctx._bufs[1], + grad_input, grad_weight, grad_bias, grad_offset, grad_mask, + grad_output, weight.shape[2], weight.shape[3], ctx.stride, + ctx.stride, ctx.padding, ctx.padding, ctx.dilation, ctx.dilation, + ctx.groups, ctx.deformable_groups, ctx.with_bias) + if not ctx.with_bias: + grad_bias = None + + return (grad_input, grad_offset, grad_mask, grad_weight, grad_bias, None, None, None, None, None) + + @staticmethod + def _infer_shape(ctx, input, weight): + n = input.size(0) + channels_out = weight.size(0) + height, width = input.shape[2:4] + kernel_h, kernel_w = weight.shape[2:4] + height_out = (height + 2 * ctx.padding - (ctx.dilation * (kernel_h - 1) + 1)) // ctx.stride + 1 + width_out = (width + 2 * ctx.padding - (ctx.dilation * (kernel_w - 1) + 1)) // ctx.stride + 1 + return n, channels_out, height_out, width_out + + +deform_conv = DeformConvFunction.apply +modulated_deform_conv = ModulatedDeformConvFunction.apply + + +class DeformConv(nn.Module): + + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + bias=False): + super(DeformConv, self).__init__() + + assert not bias + assert in_channels % groups == 0, f'in_channels {in_channels} is not divisible by groups {groups}' + assert out_channels % groups == 0, f'out_channels {out_channels} is not divisible by groups {groups}' + + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = _pair(stride) + self.padding = _pair(padding) + self.dilation = _pair(dilation) + self.groups = groups + self.deformable_groups = deformable_groups + # enable compatibility with nn.Conv2d + self.transposed = False + self.output_padding = _single(0) + + self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels // self.groups, *self.kernel_size)) + + self.reset_parameters() + + def reset_parameters(self): + n = self.in_channels + for k in self.kernel_size: + n *= k + stdv = 1. / math.sqrt(n) + self.weight.data.uniform_(-stdv, stdv) + + def forward(self, x, offset): + # To fix an assert error in deform_conv_cuda.cpp:128 + # input image is smaller than kernel + input_pad = (x.size(2) < self.kernel_size[0] or x.size(3) < self.kernel_size[1]) + if input_pad: + pad_h = max(self.kernel_size[0] - x.size(2), 0) + pad_w = max(self.kernel_size[1] - x.size(3), 0) + x = F.pad(x, (0, pad_w, 0, pad_h), 'constant', 0).contiguous() + offset = F.pad(offset, (0, pad_w, 0, pad_h), 'constant', 0).contiguous() + out = deform_conv(x, offset, self.weight, self.stride, self.padding, self.dilation, self.groups, + self.deformable_groups) + if input_pad: + out = out[:, :, :out.size(2) - pad_h, :out.size(3) - pad_w].contiguous() + return out + + +class DeformConvPack(DeformConv): + """A Deformable Conv Encapsulation that acts as normal Conv layers. + + Args: + in_channels (int): Same as nn.Conv2d. + out_channels (int): Same as nn.Conv2d. + kernel_size (int or tuple[int]): Same as nn.Conv2d. + stride (int or tuple[int]): Same as nn.Conv2d. + padding (int or tuple[int]): Same as nn.Conv2d. + dilation (int or tuple[int]): Same as nn.Conv2d. + groups (int): Same as nn.Conv2d. + bias (bool or str): If specified as `auto`, it will be decided by the + norm_cfg. Bias will be set as True if norm_cfg is None, otherwise + False. + """ + + _version = 2 + + def __init__(self, *args, **kwargs): + super(DeformConvPack, self).__init__(*args, **kwargs) + + self.conv_offset = nn.Conv2d( + self.in_channels, + self.deformable_groups * 2 * self.kernel_size[0] * self.kernel_size[1], + kernel_size=self.kernel_size, + stride=_pair(self.stride), + padding=_pair(self.padding), + dilation=_pair(self.dilation), + bias=True) + self.init_offset() + + def init_offset(self): + self.conv_offset.weight.data.zero_() + self.conv_offset.bias.data.zero_() + + def forward(self, x): + offset = self.conv_offset(x) + return deform_conv(x, offset, self.weight, self.stride, self.padding, self.dilation, self.groups, + self.deformable_groups) + + +class ModulatedDeformConv(nn.Module): + + def __init__(self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + bias=True): + super(ModulatedDeformConv, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = stride + self.padding = padding + self.dilation = dilation + self.groups = groups + self.deformable_groups = deformable_groups + self.with_bias = bias + # enable compatibility with nn.Conv2d + self.transposed = False + self.output_padding = _single(0) + + self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels // groups, *self.kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(out_channels)) + else: + self.register_parameter('bias', None) + self.init_weights() + + def init_weights(self): + n = self.in_channels + for k in self.kernel_size: + n *= k + stdv = 1. / math.sqrt(n) + self.weight.data.uniform_(-stdv, stdv) + if self.bias is not None: + self.bias.data.zero_() + + def forward(self, x, offset, mask): + return modulated_deform_conv(x, offset, mask, self.weight, self.bias, self.stride, self.padding, self.dilation, + self.groups, self.deformable_groups) + + +class ModulatedDeformConvPack(ModulatedDeformConv): + """A ModulatedDeformable Conv Encapsulation that acts as normal Conv layers. + + Args: + in_channels (int): Same as nn.Conv2d. + out_channels (int): Same as nn.Conv2d. + kernel_size (int or tuple[int]): Same as nn.Conv2d. + stride (int or tuple[int]): Same as nn.Conv2d. + padding (int or tuple[int]): Same as nn.Conv2d. + dilation (int or tuple[int]): Same as nn.Conv2d. + groups (int): Same as nn.Conv2d. + bias (bool or str): If specified as `auto`, it will be decided by the + norm_cfg. Bias will be set as True if norm_cfg is None, otherwise + False. + """ + + _version = 2 + + def __init__(self, *args, **kwargs): + super(ModulatedDeformConvPack, self).__init__(*args, **kwargs) + + self.conv_offset = nn.Conv2d( + self.in_channels, + self.deformable_groups * 3 * self.kernel_size[0] * self.kernel_size[1], + kernel_size=self.kernel_size, + stride=_pair(self.stride), + padding=_pair(self.padding), + dilation=_pair(self.dilation), + bias=True) + self.init_weights() + + def init_weights(self): + super(ModulatedDeformConvPack, self).init_weights() + if hasattr(self, 'conv_offset'): + self.conv_offset.weight.data.zero_() + self.conv_offset.bias.data.zero_() + + def forward(self, x): + out = self.conv_offset(x) + o1, o2, mask = torch.chunk(out, 3, dim=1) + offset = torch.cat((o1, o2), dim=1) + mask = torch.sigmoid(mask) + return modulated_deform_conv(x, offset, mask, self.weight, self.bias, self.stride, self.padding, self.dilation, + self.groups, self.deformable_groups) diff --git a/r_basicsr/ops/dcn/src/deform_conv_cuda.cpp b/r_basicsr/ops/dcn/src/deform_conv_cuda.cpp new file mode 100644 index 0000000..191298a --- /dev/null +++ b/r_basicsr/ops/dcn/src/deform_conv_cuda.cpp @@ -0,0 +1,685 @@ +// modify from +// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda.c + +#include +#include + +#include +#include + +void deformable_im2col(const at::Tensor data_im, const at::Tensor data_offset, + const int channels, const int height, const int width, + const int ksize_h, const int ksize_w, const int pad_h, + const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int parallel_imgs, const int deformable_group, + at::Tensor data_col); + +void deformable_col2im(const at::Tensor data_col, const at::Tensor data_offset, + const int channels, const int height, const int width, + const int ksize_h, const int ksize_w, const int pad_h, + const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int parallel_imgs, const int deformable_group, + at::Tensor grad_im); + +void deformable_col2im_coord( + const at::Tensor data_col, const at::Tensor data_im, + const at::Tensor data_offset, const int channels, const int height, + const int width, const int ksize_h, const int ksize_w, const int pad_h, + const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, const int parallel_imgs, + const int deformable_group, at::Tensor grad_offset); + +void modulated_deformable_im2col_cuda( + const at::Tensor data_im, const at::Tensor data_offset, + const at::Tensor data_mask, const int batch_size, const int channels, + const int height_im, const int width_im, const int height_col, + const int width_col, const int kernel_h, const int kenerl_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, const int deformable_group, + at::Tensor data_col); + +void modulated_deformable_col2im_cuda( + const at::Tensor data_col, const at::Tensor data_offset, + const at::Tensor data_mask, const int batch_size, const int channels, + const int height_im, const int width_im, const int height_col, + const int width_col, const int kernel_h, const int kenerl_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, const int deformable_group, + at::Tensor grad_im); + +void modulated_deformable_col2im_coord_cuda( + const at::Tensor data_col, const at::Tensor data_im, + const at::Tensor data_offset, const at::Tensor data_mask, + const int batch_size, const int channels, const int height_im, + const int width_im, const int height_col, const int width_col, + const int kernel_h, const int kenerl_w, const int pad_h, const int pad_w, + const int stride_h, const int stride_w, const int dilation_h, + const int dilation_w, const int deformable_group, at::Tensor grad_offset, + at::Tensor grad_mask); + +void shape_check(at::Tensor input, at::Tensor offset, at::Tensor *gradOutput, + at::Tensor weight, int kH, int kW, int dH, int dW, int padH, + int padW, int dilationH, int dilationW, int group, + int deformable_group) { + TORCH_CHECK(weight.ndimension() == 4, + "4D weight tensor (nOutputPlane,nInputPlane,kH,kW) expected, " + "but got: %s", + weight.ndimension()); + + TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous"); + + TORCH_CHECK(kW > 0 && kH > 0, + "kernel size should be greater than zero, but got kH: %d kW: %d", kH, + kW); + + TORCH_CHECK((weight.size(2) == kH && weight.size(3) == kW), + "kernel size should be consistent with weight, ", + "but got kH: %d kW: %d weight.size(2): %d, weight.size(3): %d", kH, + kW, weight.size(2), weight.size(3)); + + TORCH_CHECK(dW > 0 && dH > 0, + "stride should be greater than zero, but got dH: %d dW: %d", dH, dW); + + TORCH_CHECK( + dilationW > 0 && dilationH > 0, + "dilation should be greater than 0, but got dilationH: %d dilationW: %d", + dilationH, dilationW); + + int ndim = input.ndimension(); + int dimf = 0; + int dimh = 1; + int dimw = 2; + + if (ndim == 4) { + dimf++; + dimh++; + dimw++; + } + + TORCH_CHECK(ndim == 3 || ndim == 4, "3D or 4D input tensor expected but got: %s", + ndim); + + long nInputPlane = weight.size(1) * group; + long inputHeight = input.size(dimh); + long inputWidth = input.size(dimw); + long nOutputPlane = weight.size(0); + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + + TORCH_CHECK(nInputPlane % deformable_group == 0, + "input channels must divide deformable group size"); + + if (outputWidth < 1 || outputHeight < 1) + AT_ERROR( + "Given input size: (%ld x %ld x %ld). " + "Calculated output size: (%ld x %ld x %ld). Output size is too small", + nInputPlane, inputHeight, inputWidth, nOutputPlane, outputHeight, + outputWidth); + + TORCH_CHECK(input.size(1) == nInputPlane, + "invalid number of input planes, expected: %d, but got: %d", + nInputPlane, input.size(1)); + + TORCH_CHECK((inputHeight >= kH && inputWidth >= kW), + "input image is smaller than kernel"); + + TORCH_CHECK((offset.size(2) == outputHeight && offset.size(3) == outputWidth), + "invalid spatial size of offset, expected height: %d width: %d, but " + "got height: %d width: %d", + outputHeight, outputWidth, offset.size(2), offset.size(3)); + + TORCH_CHECK((offset.size(1) == deformable_group * 2 * kH * kW), + "invalid number of channels of offset"); + + if (gradOutput != NULL) { + TORCH_CHECK(gradOutput->size(dimf) == nOutputPlane, + "invalid number of gradOutput planes, expected: %d, but got: %d", + nOutputPlane, gradOutput->size(dimf)); + + TORCH_CHECK((gradOutput->size(dimh) == outputHeight && + gradOutput->size(dimw) == outputWidth), + "invalid size of gradOutput, expected height: %d width: %d , but " + "got height: %d width: %d", + outputHeight, outputWidth, gradOutput->size(dimh), + gradOutput->size(dimw)); + } +} + +int deform_conv_forward_cuda(at::Tensor input, at::Tensor weight, + at::Tensor offset, at::Tensor output, + at::Tensor columns, at::Tensor ones, int kW, + int kH, int dW, int dH, int padW, int padH, + int dilationW, int dilationH, int group, + int deformable_group, int im2col_step) { + // todo: resize columns to include im2col: done + // todo: add im2col_step as input + // todo: add new output buffer and transpose it to output (or directly + // transpose output) todo: possibly change data indexing because of + // parallel_imgs + + shape_check(input, offset, NULL, weight, kH, kW, dH, dW, padH, padW, + dilationH, dilationW, group, deformable_group); + at::DeviceGuard guard(input.device()); + + input = input.contiguous(); + offset = offset.contiguous(); + weight = weight.contiguous(); + + int batch = 1; + if (input.ndimension() == 3) { + // Force batch + batch = 0; + input.unsqueeze_(0); + offset.unsqueeze_(0); + } + + // todo: assert batchsize dividable by im2col_step + + long batchSize = input.size(0); + long nInputPlane = input.size(1); + long inputHeight = input.size(2); + long inputWidth = input.size(3); + + long nOutputPlane = weight.size(0); + + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + + TORCH_CHECK((offset.size(0) == batchSize), "invalid batch size of offset"); + + output = output.view({batchSize / im2col_step, im2col_step, nOutputPlane, + outputHeight, outputWidth}); + columns = at::zeros( + {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth}, + input.options()); + + if (ones.ndimension() != 2 || + ones.size(0) * ones.size(1) < outputHeight * outputWidth) { + ones = at::ones({outputHeight, outputWidth}, input.options()); + } + + input = input.view({batchSize / im2col_step, im2col_step, nInputPlane, + inputHeight, inputWidth}); + offset = + offset.view({batchSize / im2col_step, im2col_step, + deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + at::Tensor output_buffer = + at::zeros({batchSize / im2col_step, nOutputPlane, + im2col_step * outputHeight, outputWidth}, + output.options()); + + output_buffer = output_buffer.view( + {output_buffer.size(0), group, output_buffer.size(1) / group, + output_buffer.size(2), output_buffer.size(3)}); + + for (int elt = 0; elt < batchSize / im2col_step; elt++) { + deformable_im2col(input[elt], offset[elt], nInputPlane, inputHeight, + inputWidth, kH, kW, padH, padW, dH, dW, dilationH, + dilationW, im2col_step, deformable_group, columns); + + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + weight = weight.view({group, weight.size(0) / group, weight.size(1), + weight.size(2), weight.size(3)}); + + for (int g = 0; g < group; g++) { + output_buffer[elt][g] = output_buffer[elt][g] + .flatten(1) + .addmm_(weight[g].flatten(1), columns[g]) + .view_as(output_buffer[elt][g]); + } + } + + output_buffer = output_buffer.view( + {output_buffer.size(0), output_buffer.size(1) * output_buffer.size(2), + output_buffer.size(3), output_buffer.size(4)}); + + output_buffer = output_buffer.view({batchSize / im2col_step, nOutputPlane, + im2col_step, outputHeight, outputWidth}); + output_buffer.transpose_(1, 2); + output.copy_(output_buffer); + output = output.view({batchSize, nOutputPlane, outputHeight, outputWidth}); + + input = input.view({batchSize, nInputPlane, inputHeight, inputWidth}); + offset = offset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + if (batch == 0) { + output = output.view({nOutputPlane, outputHeight, outputWidth}); + input = input.view({nInputPlane, inputHeight, inputWidth}); + offset = offset.view({offset.size(1), offset.size(2), offset.size(3)}); + } + + return 1; +} + +int deform_conv_backward_input_cuda(at::Tensor input, at::Tensor offset, + at::Tensor gradOutput, at::Tensor gradInput, + at::Tensor gradOffset, at::Tensor weight, + at::Tensor columns, int kW, int kH, int dW, + int dH, int padW, int padH, int dilationW, + int dilationH, int group, + int deformable_group, int im2col_step) { + shape_check(input, offset, &gradOutput, weight, kH, kW, dH, dW, padH, padW, + dilationH, dilationW, group, deformable_group); + at::DeviceGuard guard(input.device()); + + input = input.contiguous(); + offset = offset.contiguous(); + gradOutput = gradOutput.contiguous(); + weight = weight.contiguous(); + + int batch = 1; + + if (input.ndimension() == 3) { + // Force batch + batch = 0; + input = input.view({1, input.size(0), input.size(1), input.size(2)}); + offset = offset.view({1, offset.size(0), offset.size(1), offset.size(2)}); + gradOutput = gradOutput.view( + {1, gradOutput.size(0), gradOutput.size(1), gradOutput.size(2)}); + } + + long batchSize = input.size(0); + long nInputPlane = input.size(1); + long inputHeight = input.size(2); + long inputWidth = input.size(3); + + long nOutputPlane = weight.size(0); + + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + + TORCH_CHECK((offset.size(0) == batchSize), 3, "invalid batch size of offset"); + gradInput = gradInput.view({batchSize, nInputPlane, inputHeight, inputWidth}); + columns = at::zeros( + {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth}, + input.options()); + + // change order of grad output + gradOutput = gradOutput.view({batchSize / im2col_step, im2col_step, + nOutputPlane, outputHeight, outputWidth}); + gradOutput.transpose_(1, 2); + + gradInput = gradInput.view({batchSize / im2col_step, im2col_step, nInputPlane, + inputHeight, inputWidth}); + input = input.view({batchSize / im2col_step, im2col_step, nInputPlane, + inputHeight, inputWidth}); + gradOffset = gradOffset.view({batchSize / im2col_step, im2col_step, + deformable_group * 2 * kH * kW, outputHeight, + outputWidth}); + offset = + offset.view({batchSize / im2col_step, im2col_step, + deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + for (int elt = 0; elt < batchSize / im2col_step; elt++) { + // divide into groups + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + weight = weight.view({group, weight.size(0) / group, weight.size(1), + weight.size(2), weight.size(3)}); + gradOutput = gradOutput.view( + {gradOutput.size(0), group, gradOutput.size(1) / group, + gradOutput.size(2), gradOutput.size(3), gradOutput.size(4)}); + + for (int g = 0; g < group; g++) { + columns[g] = columns[g].addmm_(weight[g].flatten(1).transpose(0, 1), + gradOutput[elt][g].flatten(1), 0.0f, 1.0f); + } + + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + gradOutput = gradOutput.view( + {gradOutput.size(0), gradOutput.size(1) * gradOutput.size(2), + gradOutput.size(3), gradOutput.size(4), gradOutput.size(5)}); + + deformable_col2im_coord(columns, input[elt], offset[elt], nInputPlane, + inputHeight, inputWidth, kH, kW, padH, padW, dH, dW, + dilationH, dilationW, im2col_step, deformable_group, + gradOffset[elt]); + + deformable_col2im(columns, offset[elt], nInputPlane, inputHeight, + inputWidth, kH, kW, padH, padW, dH, dW, dilationH, + dilationW, im2col_step, deformable_group, gradInput[elt]); + } + + gradOutput.transpose_(1, 2); + gradOutput = + gradOutput.view({batchSize, nOutputPlane, outputHeight, outputWidth}); + + gradInput = gradInput.view({batchSize, nInputPlane, inputHeight, inputWidth}); + input = input.view({batchSize, nInputPlane, inputHeight, inputWidth}); + gradOffset = gradOffset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + offset = offset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + if (batch == 0) { + gradOutput = gradOutput.view({nOutputPlane, outputHeight, outputWidth}); + input = input.view({nInputPlane, inputHeight, inputWidth}); + gradInput = gradInput.view({nInputPlane, inputHeight, inputWidth}); + offset = offset.view({offset.size(1), offset.size(2), offset.size(3)}); + gradOffset = + gradOffset.view({offset.size(1), offset.size(2), offset.size(3)}); + } + + return 1; +} + +int deform_conv_backward_parameters_cuda( + at::Tensor input, at::Tensor offset, at::Tensor gradOutput, + at::Tensor gradWeight, // at::Tensor gradBias, + at::Tensor columns, at::Tensor ones, int kW, int kH, int dW, int dH, + int padW, int padH, int dilationW, int dilationH, int group, + int deformable_group, float scale, int im2col_step) { + // todo: transpose and reshape outGrad + // todo: reshape columns + // todo: add im2col_step as input + + shape_check(input, offset, &gradOutput, gradWeight, kH, kW, dH, dW, padH, + padW, dilationH, dilationW, group, deformable_group); + at::DeviceGuard guard(input.device()); + + input = input.contiguous(); + offset = offset.contiguous(); + gradOutput = gradOutput.contiguous(); + + int batch = 1; + + if (input.ndimension() == 3) { + // Force batch + batch = 0; + input = input.view( + at::IntList({1, input.size(0), input.size(1), input.size(2)})); + gradOutput = gradOutput.view( + {1, gradOutput.size(0), gradOutput.size(1), gradOutput.size(2)}); + } + + long batchSize = input.size(0); + long nInputPlane = input.size(1); + long inputHeight = input.size(2); + long inputWidth = input.size(3); + + long nOutputPlane = gradWeight.size(0); + + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + + TORCH_CHECK((offset.size(0) == batchSize), "invalid batch size of offset"); + + columns = at::zeros( + {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth}, + input.options()); + + gradOutput = gradOutput.view({batchSize / im2col_step, im2col_step, + nOutputPlane, outputHeight, outputWidth}); + gradOutput.transpose_(1, 2); + + at::Tensor gradOutputBuffer = at::zeros_like(gradOutput); + gradOutputBuffer = + gradOutputBuffer.view({batchSize / im2col_step, nOutputPlane, im2col_step, + outputHeight, outputWidth}); + gradOutputBuffer.copy_(gradOutput); + gradOutputBuffer = + gradOutputBuffer.view({batchSize / im2col_step, nOutputPlane, + im2col_step * outputHeight, outputWidth}); + + gradOutput.transpose_(1, 2); + gradOutput = + gradOutput.view({batchSize, nOutputPlane, outputHeight, outputWidth}); + + input = input.view({batchSize / im2col_step, im2col_step, nInputPlane, + inputHeight, inputWidth}); + offset = + offset.view({batchSize / im2col_step, im2col_step, + deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + for (int elt = 0; elt < batchSize / im2col_step; elt++) { + deformable_im2col(input[elt], offset[elt], nInputPlane, inputHeight, + inputWidth, kH, kW, padH, padW, dH, dW, dilationH, + dilationW, im2col_step, deformable_group, columns); + + // divide into group + gradOutputBuffer = gradOutputBuffer.view( + {gradOutputBuffer.size(0), group, gradOutputBuffer.size(1) / group, + gradOutputBuffer.size(2), gradOutputBuffer.size(3)}); + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + gradWeight = + gradWeight.view({group, gradWeight.size(0) / group, gradWeight.size(1), + gradWeight.size(2), gradWeight.size(3)}); + + for (int g = 0; g < group; g++) { + gradWeight[g] = gradWeight[g] + .flatten(1) + .addmm_(gradOutputBuffer[elt][g].flatten(1), + columns[g].transpose(1, 0), 1.0, scale) + .view_as(gradWeight[g]); + } + gradOutputBuffer = gradOutputBuffer.view( + {gradOutputBuffer.size(0), + gradOutputBuffer.size(1) * gradOutputBuffer.size(2), + gradOutputBuffer.size(3), gradOutputBuffer.size(4)}); + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + gradWeight = gradWeight.view({gradWeight.size(0) * gradWeight.size(1), + gradWeight.size(2), gradWeight.size(3), + gradWeight.size(4)}); + } + + input = input.view({batchSize, nInputPlane, inputHeight, inputWidth}); + offset = offset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + if (batch == 0) { + gradOutput = gradOutput.view({nOutputPlane, outputHeight, outputWidth}); + input = input.view({nInputPlane, inputHeight, inputWidth}); + } + + return 1; +} + +void modulated_deform_conv_cuda_forward( + at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones, + at::Tensor offset, at::Tensor mask, at::Tensor output, at::Tensor columns, + int kernel_h, int kernel_w, const int stride_h, const int stride_w, + const int pad_h, const int pad_w, const int dilation_h, + const int dilation_w, const int group, const int deformable_group, + const bool with_bias) { + TORCH_CHECK(input.is_contiguous(), "input tensor has to be contiguous"); + TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous"); + at::DeviceGuard guard(input.device()); + + const int batch = input.size(0); + const int channels = input.size(1); + const int height = input.size(2); + const int width = input.size(3); + + const int channels_out = weight.size(0); + const int channels_kernel = weight.size(1); + const int kernel_h_ = weight.size(2); + const int kernel_w_ = weight.size(3); + + if (kernel_h_ != kernel_h || kernel_w_ != kernel_w) + AT_ERROR("Input shape and kernel shape won't match: (%d x %d vs %d x %d).", + kernel_h_, kernel_w, kernel_h_, kernel_w_); + if (channels != channels_kernel * group) + AT_ERROR("Input shape and kernel channels won't match: (%d vs %d).", + channels, channels_kernel * group); + + const int height_out = + (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; + const int width_out = + (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1; + + if (ones.ndimension() != 2 || + ones.size(0) * ones.size(1) < height_out * width_out) { + // Resize plane and fill with ones... + ones = at::ones({height_out, width_out}, input.options()); + } + + // resize output + output = output.view({batch, channels_out, height_out, width_out}).zero_(); + // resize temporary columns + columns = + at::zeros({channels * kernel_h * kernel_w, 1 * height_out * width_out}, + input.options()); + + output = output.view({output.size(0), group, output.size(1) / group, + output.size(2), output.size(3)}); + + for (int b = 0; b < batch; b++) { + modulated_deformable_im2col_cuda( + input[b], offset[b], mask[b], 1, channels, height, width, height_out, + width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, deformable_group, columns); + + // divide into group + weight = weight.view({group, weight.size(0) / group, weight.size(1), + weight.size(2), weight.size(3)}); + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + + for (int g = 0; g < group; g++) { + output[b][g] = output[b][g] + .flatten(1) + .addmm_(weight[g].flatten(1), columns[g]) + .view_as(output[b][g]); + } + + weight = weight.view({weight.size(0) * weight.size(1), weight.size(2), + weight.size(3), weight.size(4)}); + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + } + + output = output.view({output.size(0), output.size(1) * output.size(2), + output.size(3), output.size(4)}); + + if (with_bias) { + output += bias.view({1, bias.size(0), 1, 1}); + } +} + +void modulated_deform_conv_cuda_backward( + at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones, + at::Tensor offset, at::Tensor mask, at::Tensor columns, + at::Tensor grad_input, at::Tensor grad_weight, at::Tensor grad_bias, + at::Tensor grad_offset, at::Tensor grad_mask, at::Tensor grad_output, + int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h, + int pad_w, int dilation_h, int dilation_w, int group, int deformable_group, + const bool with_bias) { + TORCH_CHECK(input.is_contiguous(), "input tensor has to be contiguous"); + TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous"); + at::DeviceGuard guard(input.device()); + + const int batch = input.size(0); + const int channels = input.size(1); + const int height = input.size(2); + const int width = input.size(3); + + const int channels_kernel = weight.size(1); + const int kernel_h_ = weight.size(2); + const int kernel_w_ = weight.size(3); + if (kernel_h_ != kernel_h || kernel_w_ != kernel_w) + AT_ERROR("Input shape and kernel shape won't match: (%d x %d vs %d x %d).", + kernel_h_, kernel_w, kernel_h_, kernel_w_); + if (channels != channels_kernel * group) + AT_ERROR("Input shape and kernel channels won't match: (%d vs %d).", + channels, channels_kernel * group); + + const int height_out = + (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; + const int width_out = + (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1; + + if (ones.ndimension() != 2 || + ones.size(0) * ones.size(1) < height_out * width_out) { + // Resize plane and fill with ones... + ones = at::ones({height_out, width_out}, input.options()); + } + + grad_input = grad_input.view({batch, channels, height, width}); + columns = at::zeros({channels * kernel_h * kernel_w, height_out * width_out}, + input.options()); + + grad_output = + grad_output.view({grad_output.size(0), group, grad_output.size(1) / group, + grad_output.size(2), grad_output.size(3)}); + + for (int b = 0; b < batch; b++) { + // divide int group + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + weight = weight.view({group, weight.size(0) / group, weight.size(1), + weight.size(2), weight.size(3)}); + + for (int g = 0; g < group; g++) { + columns[g].addmm_(weight[g].flatten(1).transpose(0, 1), + grad_output[b][g].flatten(1), 0.0f, 1.0f); + } + + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + weight = weight.view({weight.size(0) * weight.size(1), weight.size(2), + weight.size(3), weight.size(4)}); + + // gradient w.r.t. input coordinate data + modulated_deformable_col2im_coord_cuda( + columns, input[b], offset[b], mask[b], 1, channels, height, width, + height_out, width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, + stride_w, dilation_h, dilation_w, deformable_group, grad_offset[b], + grad_mask[b]); + // gradient w.r.t. input data + modulated_deformable_col2im_cuda( + columns, offset[b], mask[b], 1, channels, height, width, height_out, + width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, deformable_group, grad_input[b]); + + // gradient w.r.t. weight, dWeight should accumulate across the batch and + // group + modulated_deformable_im2col_cuda( + input[b], offset[b], mask[b], 1, channels, height, width, height_out, + width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, deformable_group, columns); + + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + grad_weight = grad_weight.view({group, grad_weight.size(0) / group, + grad_weight.size(1), grad_weight.size(2), + grad_weight.size(3)}); + if (with_bias) + grad_bias = grad_bias.view({group, grad_bias.size(0) / group}); + + for (int g = 0; g < group; g++) { + grad_weight[g] = + grad_weight[g] + .flatten(1) + .addmm_(grad_output[b][g].flatten(1), columns[g].transpose(0, 1)) + .view_as(grad_weight[g]); + if (with_bias) { + grad_bias[g] = + grad_bias[g] + .view({-1, 1}) + .addmm_(grad_output[b][g].flatten(1), ones.view({-1, 1})) + .view(-1); + } + } + + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + grad_weight = grad_weight.view({grad_weight.size(0) * grad_weight.size(1), + grad_weight.size(2), grad_weight.size(3), + grad_weight.size(4)}); + if (with_bias) + grad_bias = grad_bias.view({grad_bias.size(0) * grad_bias.size(1)}); + } + grad_output = grad_output.view({grad_output.size(0) * grad_output.size(1), + grad_output.size(2), grad_output.size(3), + grad_output.size(4)}); +} diff --git a/r_basicsr/ops/dcn/src/deform_conv_cuda_kernel.cu b/r_basicsr/ops/dcn/src/deform_conv_cuda_kernel.cu new file mode 100644 index 0000000..9fe9ba3 --- /dev/null +++ b/r_basicsr/ops/dcn/src/deform_conv_cuda_kernel.cu @@ -0,0 +1,867 @@ +/*! + ******************* BEGIN Caffe Copyright Notice and Disclaimer **************** + * + * COPYRIGHT + * + * All contributions by the University of California: + * Copyright (c) 2014-2017 The Regents of the University of California (Regents) + * All rights reserved. + * + * All other contributions: + * Copyright (c) 2014-2017, the respective contributors + * All rights reserved. + * + * Caffe uses a shared copyright model: each contributor holds copyright over + * their contributions to Caffe. The project versioning records all such + * contribution and copyright details. If a contributor wants to further mark + * their specific copyright on a particular contribution, they should indicate + * their copyright solely in the commit message of the change when it is + * committed. + * + * LICENSE + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are met: + * + * 1. Redistributions of source code must retain the above copyright notice, this + * list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright notice, + * this list of conditions and the following disclaimer in the documentation + * and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND + * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * CONTRIBUTION AGREEMENT + * + * By contributing to the BVLC/caffe repository through pull-request, comment, + * or otherwise, the contributor releases their content to the + * license and copyright terms herein. + * + ***************** END Caffe Copyright Notice and Disclaimer ******************** + * + * Copyright (c) 2018 Microsoft + * Licensed under The MIT License [see LICENSE for details] + * \file modulated_deformable_im2col.cuh + * \brief Function definitions of converting an image to + * column matrix based on kernel, padding, dilation, and offset. + * These functions are mainly used in deformable convolution operators. + * \ref: https://arxiv.org/abs/1703.06211 + * \author Yuwen Xiong, Haozhi Qi, Jifeng Dai, Xizhou Zhu, Han Hu, Dazhi Cheng + */ + +// modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda_kernel.cu + +#include +#include +#include +#include +#include +#include + +using namespace at; + +#define CUDA_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +const int CUDA_NUM_THREADS = 1024; +const int kMaxGridNum = 65535; + +inline int GET_BLOCKS(const int N) +{ + return std::min(kMaxGridNum, (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS); +} + +template +__device__ scalar_t deformable_im2col_bilinear(const scalar_t *bottom_data, const int data_width, + const int height, const int width, scalar_t h, scalar_t w) +{ + + int h_low = floor(h); + int w_low = floor(w); + int h_high = h_low + 1; + int w_high = w_low + 1; + + scalar_t lh = h - h_low; + scalar_t lw = w - w_low; + scalar_t hh = 1 - lh, hw = 1 - lw; + + scalar_t v1 = 0; + if (h_low >= 0 && w_low >= 0) + v1 = bottom_data[h_low * data_width + w_low]; + scalar_t v2 = 0; + if (h_low >= 0 && w_high <= width - 1) + v2 = bottom_data[h_low * data_width + w_high]; + scalar_t v3 = 0; + if (h_high <= height - 1 && w_low >= 0) + v3 = bottom_data[h_high * data_width + w_low]; + scalar_t v4 = 0; + if (h_high <= height - 1 && w_high <= width - 1) + v4 = bottom_data[h_high * data_width + w_high]; + + scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; + + scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +template +__device__ scalar_t get_gradient_weight(scalar_t argmax_h, scalar_t argmax_w, + const int h, const int w, const int height, const int width) +{ + + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width) + { + //empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + if (h == argmax_h_low && w == argmax_w_low) + weight = (h + 1 - argmax_h) * (w + 1 - argmax_w); + if (h == argmax_h_low && w == argmax_w_high) + weight = (h + 1 - argmax_h) * (argmax_w + 1 - w); + if (h == argmax_h_high && w == argmax_w_low) + weight = (argmax_h + 1 - h) * (w + 1 - argmax_w); + if (h == argmax_h_high && w == argmax_w_high) + weight = (argmax_h + 1 - h) * (argmax_w + 1 - w); + return weight; +} + +template +__device__ scalar_t get_coordinate_weight(scalar_t argmax_h, scalar_t argmax_w, + const int height, const int width, const scalar_t *im_data, + const int data_width, const int bp_dir) +{ + + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width) + { + //empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + + if (bp_dir == 0) + { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += -1 * (argmax_w - argmax_w_low) * im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_w - argmax_w_low) * im_data[argmax_h_high * data_width + argmax_w_high]; + } + else if (bp_dir == 1) + { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += -1 * (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_high]; + } + + return weight; +} + +template +__global__ void deformable_im2col_gpu_kernel(const int n, const scalar_t *data_im, const scalar_t *data_offset, + const int height, const int width, const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, const int channel_per_deformable_group, + const int batch_size, const int num_channels, const int deformable_group, + const int height_col, const int width_col, + scalar_t *data_col) +{ + CUDA_KERNEL_LOOP(index, n) + { + // index index of output matrix + const int w_col = index % width_col; + const int h_col = (index / width_col) % height_col; + const int b_col = (index / width_col / height_col) % batch_size; + const int c_im = (index / width_col / height_col) / batch_size; + const int c_col = c_im * kernel_h * kernel_w; + + // compute deformable group index + const int deformable_group_index = c_im / channel_per_deformable_group; + + const int h_in = h_col * stride_h - pad_h; + const int w_in = w_col * stride_w - pad_w; + scalar_t *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + //const scalar_t* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in; + const scalar_t *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width; + const scalar_t *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col; + + for (int i = 0; i < kernel_h; ++i) + { + for (int j = 0; j < kernel_w; ++j) + { + const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; + const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + scalar_t val = static_cast(0); + const scalar_t h_im = h_in + i * dilation_h + offset_h; + const scalar_t w_im = w_in + j * dilation_w + offset_w; + if (h_im > -1 && w_im > -1 && h_im < height && w_im < width) + { + //const scalar_t map_h = i * dilation_h + offset_h; + //const scalar_t map_w = j * dilation_w + offset_w; + //const int cur_height = height - h_in; + //const int cur_width = width - w_in; + //val = deformable_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w); + val = deformable_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im); + } + *data_col_ptr = val; + data_col_ptr += batch_size * height_col * width_col; + } + } + } +} + +void deformable_im2col( + const at::Tensor data_im, const at::Tensor data_offset, const int channels, + const int height, const int width, const int ksize_h, const int ksize_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, const int parallel_imgs, + const int deformable_group, at::Tensor data_col) +{ + // num_axes should be smaller than block size + // todo: check parallel_imgs is correctly passed in + int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1; + int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1; + int num_kernels = channels * height_col * width_col * parallel_imgs; + int channel_per_deformable_group = channels / deformable_group; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_im.scalar_type(), "deformable_im2col_gpu", ([&] { + const scalar_t *data_im_ = data_im.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + scalar_t *data_col_ = data_col.data_ptr(); + + deformable_im2col_gpu_kernel<<>>( + num_kernels, data_im_, data_offset_, height, width, ksize_h, ksize_w, + pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, + channel_per_deformable_group, parallel_imgs, channels, deformable_group, + height_col, width_col, data_col_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) + { + printf("error in deformable_im2col: %s\n", cudaGetErrorString(err)); + } +} + +template +__global__ void deformable_col2im_gpu_kernel( + const int n, const scalar_t *data_col, const scalar_t *data_offset, + const int channels, const int height, const int width, + const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, const int deformable_group, + const int height_col, const int width_col, + scalar_t *grad_im) +{ + CUDA_KERNEL_LOOP(index, n) + { + const int j = (index / width_col / height_col / batch_size) % kernel_w; + const int i = (index / width_col / height_col / batch_size / kernel_w) % kernel_h; + const int c = index / width_col / height_col / batch_size / kernel_w / kernel_h; + // compute the start and end of the output + + const int deformable_group_index = c / channel_per_deformable_group; + + int w_out = index % width_col; + int h_out = (index / width_col) % height_col; + int b = (index / width_col / height_col) % batch_size; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + + const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * + 2 * kernel_h * kernel_w * height_col * width_col; + const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out; + const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h; + const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w; + + const scalar_t cur_top_grad = data_col[index]; + const int cur_h = (int)cur_inv_h_data; + const int cur_w = (int)cur_inv_w_data; + for (int dy = -2; dy <= 2; dy++) + { + for (int dx = -2; dx <= 2; dx++) + { + if (cur_h + dy >= 0 && cur_h + dy < height && + cur_w + dx >= 0 && cur_w + dx < width && + abs(cur_inv_h_data - (cur_h + dy)) < 1 && + abs(cur_inv_w_data - (cur_w + dx)) < 1) + { + int cur_bottom_grad_pos = ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx; + scalar_t weight = get_gradient_weight(cur_inv_h_data, cur_inv_w_data, cur_h + dy, cur_w + dx, height, width); + atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad); + } + } + } + } +} + +void deformable_col2im( + const at::Tensor data_col, const at::Tensor data_offset, const int channels, + const int height, const int width, const int ksize_h, + const int ksize_w, const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int parallel_imgs, const int deformable_group, + at::Tensor grad_im) +{ + + // todo: make sure parallel_imgs is passed in correctly + int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1; + int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1; + int num_kernels = channels * ksize_h * ksize_w * height_col * width_col * parallel_imgs; + int channel_per_deformable_group = channels / deformable_group; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "deformable_col2im_gpu", ([&] { + const scalar_t *data_col_ = data_col.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + scalar_t *grad_im_ = grad_im.data_ptr(); + + deformable_col2im_gpu_kernel<<>>( + num_kernels, data_col_, data_offset_, channels, height, width, ksize_h, + ksize_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, channel_per_deformable_group, + parallel_imgs, deformable_group, height_col, width_col, grad_im_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) + { + printf("error in deformable_col2im: %s\n", cudaGetErrorString(err)); + } +} + +template +__global__ void deformable_col2im_coord_gpu_kernel(const int n, const scalar_t *data_col, + const scalar_t *data_im, const scalar_t *data_offset, + const int channels, const int height, const int width, + const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, const int offset_channels, const int deformable_group, + const int height_col, const int width_col, scalar_t *grad_offset) +{ + CUDA_KERNEL_LOOP(index, n) + { + scalar_t val = 0; + int w = index % width_col; + int h = (index / width_col) % height_col; + int c = (index / width_col / height_col) % offset_channels; + int b = (index / width_col / height_col) / offset_channels; + // compute the start and end of the output + + const int deformable_group_index = c / (2 * kernel_h * kernel_w); + const int col_step = kernel_h * kernel_w; + int cnt = 0; + const scalar_t *data_col_ptr = data_col + deformable_group_index * channel_per_deformable_group * + batch_size * width_col * height_col; + const scalar_t *data_im_ptr = data_im + (b * deformable_group + deformable_group_index) * + channel_per_deformable_group / kernel_h / kernel_w * height * width; + const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * + kernel_h * kernel_w * height_col * width_col; + + const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w; + + for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; col_c += col_step) + { + const int col_pos = (((col_c * batch_size + b) * height_col) + h) * width_col + w; + const int bp_dir = offset_c % 2; + + int j = (col_pos / width_col / height_col / batch_size) % kernel_w; + int i = (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h; + int w_out = col_pos % width_col; + int h_out = (col_pos / width_col) % height_col; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + const int data_offset_h_ptr = (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out); + const int data_offset_w_ptr = (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out); + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + scalar_t inv_h = h_in + i * dilation_h + offset_h; + scalar_t inv_w = w_in + j * dilation_w + offset_w; + if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width) + { + inv_h = inv_w = -2; + } + const scalar_t weight = get_coordinate_weight( + inv_h, inv_w, + height, width, data_im_ptr + cnt * height * width, width, bp_dir); + val += weight * data_col_ptr[col_pos]; + cnt += 1; + } + + grad_offset[index] = val; + } +} + +void deformable_col2im_coord( + const at::Tensor data_col, const at::Tensor data_im, const at::Tensor data_offset, + const int channels, const int height, const int width, const int ksize_h, + const int ksize_w, const int pad_h, const int pad_w, const int stride_h, + const int stride_w, const int dilation_h, const int dilation_w, + const int parallel_imgs, const int deformable_group, at::Tensor grad_offset) +{ + + int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1; + int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1; + int num_kernels = height_col * width_col * 2 * ksize_h * ksize_w * deformable_group * parallel_imgs; + int channel_per_deformable_group = channels * ksize_h * ksize_w / deformable_group; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "deformable_col2im_coord_gpu", ([&] { + const scalar_t *data_col_ = data_col.data_ptr(); + const scalar_t *data_im_ = data_im.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + scalar_t *grad_offset_ = grad_offset.data_ptr(); + + deformable_col2im_coord_gpu_kernel<<>>( + num_kernels, data_col_, data_im_, data_offset_, channels, height, width, + ksize_h, ksize_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, channel_per_deformable_group, + parallel_imgs, 2 * ksize_h * ksize_w * deformable_group, deformable_group, + height_col, width_col, grad_offset_); + })); +} + +template +__device__ scalar_t dmcn_im2col_bilinear(const scalar_t *bottom_data, const int data_width, + const int height, const int width, scalar_t h, scalar_t w) +{ + int h_low = floor(h); + int w_low = floor(w); + int h_high = h_low + 1; + int w_high = w_low + 1; + + scalar_t lh = h - h_low; + scalar_t lw = w - w_low; + scalar_t hh = 1 - lh, hw = 1 - lw; + + scalar_t v1 = 0; + if (h_low >= 0 && w_low >= 0) + v1 = bottom_data[h_low * data_width + w_low]; + scalar_t v2 = 0; + if (h_low >= 0 && w_high <= width - 1) + v2 = bottom_data[h_low * data_width + w_high]; + scalar_t v3 = 0; + if (h_high <= height - 1 && w_low >= 0) + v3 = bottom_data[h_high * data_width + w_low]; + scalar_t v4 = 0; + if (h_high <= height - 1 && w_high <= width - 1) + v4 = bottom_data[h_high * data_width + w_high]; + + scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; + + scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +template +__device__ scalar_t dmcn_get_gradient_weight(scalar_t argmax_h, scalar_t argmax_w, + const int h, const int w, const int height, const int width) +{ + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width) + { + //empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + if (h == argmax_h_low && w == argmax_w_low) + weight = (h + 1 - argmax_h) * (w + 1 - argmax_w); + if (h == argmax_h_low && w == argmax_w_high) + weight = (h + 1 - argmax_h) * (argmax_w + 1 - w); + if (h == argmax_h_high && w == argmax_w_low) + weight = (argmax_h + 1 - h) * (w + 1 - argmax_w); + if (h == argmax_h_high && w == argmax_w_high) + weight = (argmax_h + 1 - h) * (argmax_w + 1 - w); + return weight; +} + +template +__device__ scalar_t dmcn_get_coordinate_weight(scalar_t argmax_h, scalar_t argmax_w, + const int height, const int width, const scalar_t *im_data, + const int data_width, const int bp_dir) +{ + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width) + { + //empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + + if (bp_dir == 0) + { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += -1 * (argmax_w - argmax_w_low) * im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_w - argmax_w_low) * im_data[argmax_h_high * data_width + argmax_w_high]; + } + else if (bp_dir == 1) + { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += -1 * (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_high]; + } + + return weight; +} + +template +__global__ void modulated_deformable_im2col_gpu_kernel(const int n, + const scalar_t *data_im, const scalar_t *data_offset, const scalar_t *data_mask, + const int height, const int width, const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, const int num_channels, const int deformable_group, + const int height_col, const int width_col, + scalar_t *data_col) +{ + CUDA_KERNEL_LOOP(index, n) + { + // index index of output matrix + const int w_col = index % width_col; + const int h_col = (index / width_col) % height_col; + const int b_col = (index / width_col / height_col) % batch_size; + const int c_im = (index / width_col / height_col) / batch_size; + const int c_col = c_im * kernel_h * kernel_w; + + // compute deformable group index + const int deformable_group_index = c_im / channel_per_deformable_group; + + const int h_in = h_col * stride_h - pad_h; + const int w_in = w_col * stride_w - pad_w; + + scalar_t *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + //const float* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in; + const scalar_t *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width; + const scalar_t *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col; + + const scalar_t *data_mask_ptr = data_mask + (b_col * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col; + + for (int i = 0; i < kernel_h; ++i) + { + for (int j = 0; j < kernel_w; ++j) + { + const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; + const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col; + const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_col) * width_col + w_col; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t mask = data_mask_ptr[data_mask_hw_ptr]; + scalar_t val = static_cast(0); + const scalar_t h_im = h_in + i * dilation_h + offset_h; + const scalar_t w_im = w_in + j * dilation_w + offset_w; + //if (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) { + if (h_im > -1 && w_im > -1 && h_im < height && w_im < width) + { + //const float map_h = i * dilation_h + offset_h; + //const float map_w = j * dilation_w + offset_w; + //const int cur_height = height - h_in; + //const int cur_width = width - w_in; + //val = dmcn_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w); + val = dmcn_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im); + } + *data_col_ptr = val * mask; + data_col_ptr += batch_size * height_col * width_col; + //data_col_ptr += height_col * width_col; + } + } + } +} + +template +__global__ void modulated_deformable_col2im_gpu_kernel(const int n, + const scalar_t *data_col, const scalar_t *data_offset, const scalar_t *data_mask, + const int channels, const int height, const int width, + const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, const int deformable_group, + const int height_col, const int width_col, + scalar_t *grad_im) +{ + CUDA_KERNEL_LOOP(index, n) + { + const int j = (index / width_col / height_col / batch_size) % kernel_w; + const int i = (index / width_col / height_col / batch_size / kernel_w) % kernel_h; + const int c = index / width_col / height_col / batch_size / kernel_w / kernel_h; + // compute the start and end of the output + + const int deformable_group_index = c / channel_per_deformable_group; + + int w_out = index % width_col; + int h_out = (index / width_col) % height_col; + int b = (index / width_col / height_col) % batch_size; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + + const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col; + const scalar_t *data_mask_ptr = data_mask + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col; + const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out; + const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out; + const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_out) * width_col + w_out; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t mask = data_mask_ptr[data_mask_hw_ptr]; + const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h; + const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w; + + const scalar_t cur_top_grad = data_col[index] * mask; + const int cur_h = (int)cur_inv_h_data; + const int cur_w = (int)cur_inv_w_data; + for (int dy = -2; dy <= 2; dy++) + { + for (int dx = -2; dx <= 2; dx++) + { + if (cur_h + dy >= 0 && cur_h + dy < height && + cur_w + dx >= 0 && cur_w + dx < width && + abs(cur_inv_h_data - (cur_h + dy)) < 1 && + abs(cur_inv_w_data - (cur_w + dx)) < 1) + { + int cur_bottom_grad_pos = ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx; + scalar_t weight = dmcn_get_gradient_weight(cur_inv_h_data, cur_inv_w_data, cur_h + dy, cur_w + dx, height, width); + atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad); + } + } + } + } +} + +template +__global__ void modulated_deformable_col2im_coord_gpu_kernel(const int n, + const scalar_t *data_col, const scalar_t *data_im, + const scalar_t *data_offset, const scalar_t *data_mask, + const int channels, const int height, const int width, + const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, const int offset_channels, const int deformable_group, + const int height_col, const int width_col, + scalar_t *grad_offset, scalar_t *grad_mask) +{ + CUDA_KERNEL_LOOP(index, n) + { + scalar_t val = 0, mval = 0; + int w = index % width_col; + int h = (index / width_col) % height_col; + int c = (index / width_col / height_col) % offset_channels; + int b = (index / width_col / height_col) / offset_channels; + // compute the start and end of the output + + const int deformable_group_index = c / (2 * kernel_h * kernel_w); + const int col_step = kernel_h * kernel_w; + int cnt = 0; + const scalar_t *data_col_ptr = data_col + deformable_group_index * channel_per_deformable_group * batch_size * width_col * height_col; + const scalar_t *data_im_ptr = data_im + (b * deformable_group + deformable_group_index) * channel_per_deformable_group / kernel_h / kernel_w * height * width; + const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col; + const scalar_t *data_mask_ptr = data_mask + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col; + + const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w; + + for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; col_c += col_step) + { + const int col_pos = (((col_c * batch_size + b) * height_col) + h) * width_col + w; + const int bp_dir = offset_c % 2; + + int j = (col_pos / width_col / height_col / batch_size) % kernel_w; + int i = (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h; + int w_out = col_pos % width_col; + int h_out = (col_pos / width_col) % height_col; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + const int data_offset_h_ptr = (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out); + const int data_offset_w_ptr = (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out); + const int data_mask_hw_ptr = (((i * kernel_w + j) * height_col + h_out) * width_col + w_out); + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t mask = data_mask_ptr[data_mask_hw_ptr]; + scalar_t inv_h = h_in + i * dilation_h + offset_h; + scalar_t inv_w = w_in + j * dilation_w + offset_w; + if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width) + { + inv_h = inv_w = -2; + } + else + { + mval += data_col_ptr[col_pos] * dmcn_im2col_bilinear(data_im_ptr + cnt * height * width, width, height, width, inv_h, inv_w); + } + const scalar_t weight = dmcn_get_coordinate_weight( + inv_h, inv_w, + height, width, data_im_ptr + cnt * height * width, width, bp_dir); + val += weight * data_col_ptr[col_pos] * mask; + cnt += 1; + } + // KERNEL_ASSIGN(grad_offset[index], offset_req, val); + grad_offset[index] = val; + if (offset_c % 2 == 0) + // KERNEL_ASSIGN(grad_mask[(((b * deformable_group + deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * height_col + h) * width_col + w], mask_req, mval); + grad_mask[(((b * deformable_group + deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * height_col + h) * width_col + w] = mval; + } +} + +void modulated_deformable_im2col_cuda( + const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask, + const int batch_size, const int channels, const int height_im, const int width_im, + const int height_col, const int width_col, const int kernel_h, const int kenerl_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int deformable_group, at::Tensor data_col) +{ + // num_axes should be smaller than block size + const int channel_per_deformable_group = channels / deformable_group; + const int num_kernels = channels * batch_size * height_col * width_col; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_im.scalar_type(), "modulated_deformable_im2col_gpu", ([&] { + const scalar_t *data_im_ = data_im.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + const scalar_t *data_mask_ = data_mask.data_ptr(); + scalar_t *data_col_ = data_col.data_ptr(); + + modulated_deformable_im2col_gpu_kernel<<>>( + num_kernels, data_im_, data_offset_, data_mask_, height_im, width_im, kernel_h, kenerl_w, + pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, channel_per_deformable_group, + batch_size, channels, deformable_group, height_col, width_col, data_col_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) + { + printf("error in modulated_deformable_im2col_cuda: %s\n", cudaGetErrorString(err)); + } +} + +void modulated_deformable_col2im_cuda( + const at::Tensor data_col, const at::Tensor data_offset, const at::Tensor data_mask, + const int batch_size, const int channels, const int height_im, const int width_im, + const int height_col, const int width_col, const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int deformable_group, at::Tensor grad_im) +{ + + const int channel_per_deformable_group = channels / deformable_group; + const int num_kernels = channels * kernel_h * kernel_w * batch_size * height_col * width_col; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "modulated_deformable_col2im_gpu", ([&] { + const scalar_t *data_col_ = data_col.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + const scalar_t *data_mask_ = data_mask.data_ptr(); + scalar_t *grad_im_ = grad_im.data_ptr(); + + modulated_deformable_col2im_gpu_kernel<<>>( + num_kernels, data_col_, data_offset_, data_mask_, channels, height_im, width_im, + kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, channel_per_deformable_group, + batch_size, deformable_group, height_col, width_col, grad_im_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) + { + printf("error in modulated_deformable_col2im_cuda: %s\n", cudaGetErrorString(err)); + } +} + +void modulated_deformable_col2im_coord_cuda( + const at::Tensor data_col, const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask, + const int batch_size, const int channels, const int height_im, const int width_im, + const int height_col, const int width_col, const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int deformable_group, + at::Tensor grad_offset, at::Tensor grad_mask) +{ + const int num_kernels = batch_size * height_col * width_col * 2 * kernel_h * kernel_w * deformable_group; + const int channel_per_deformable_group = channels * kernel_h * kernel_w / deformable_group; + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "modulated_deformable_col2im_coord_gpu", ([&] { + const scalar_t *data_col_ = data_col.data_ptr(); + const scalar_t *data_im_ = data_im.data_ptr(); + const scalar_t *data_offset_ = data_offset.data_ptr(); + const scalar_t *data_mask_ = data_mask.data_ptr(); + scalar_t *grad_offset_ = grad_offset.data_ptr(); + scalar_t *grad_mask_ = grad_mask.data_ptr(); + + modulated_deformable_col2im_coord_gpu_kernel<<>>( + num_kernels, data_col_, data_im_, data_offset_, data_mask_, channels, height_im, width_im, + kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w, + dilation_h, dilation_w, channel_per_deformable_group, + batch_size, 2 * kernel_h * kernel_w * deformable_group, deformable_group, height_col, width_col, + grad_offset_, grad_mask_); + })); + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) + { + printf("error in modulated_deformable_col2im_coord_cuda: %s\n", cudaGetErrorString(err)); + } +} diff --git a/r_basicsr/ops/dcn/src/deform_conv_ext.cpp b/r_basicsr/ops/dcn/src/deform_conv_ext.cpp new file mode 100644 index 0000000..5c21d02 --- /dev/null +++ b/r_basicsr/ops/dcn/src/deform_conv_ext.cpp @@ -0,0 +1,164 @@ +// modify from +// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda.c + +#include +#include + +#include +#include + +#define WITH_CUDA // always use cuda +#ifdef WITH_CUDA +int deform_conv_forward_cuda(at::Tensor input, at::Tensor weight, + at::Tensor offset, at::Tensor output, + at::Tensor columns, at::Tensor ones, int kW, + int kH, int dW, int dH, int padW, int padH, + int dilationW, int dilationH, int group, + int deformable_group, int im2col_step); + +int deform_conv_backward_input_cuda(at::Tensor input, at::Tensor offset, + at::Tensor gradOutput, at::Tensor gradInput, + at::Tensor gradOffset, at::Tensor weight, + at::Tensor columns, int kW, int kH, int dW, + int dH, int padW, int padH, int dilationW, + int dilationH, int group, + int deformable_group, int im2col_step); + +int deform_conv_backward_parameters_cuda( + at::Tensor input, at::Tensor offset, at::Tensor gradOutput, + at::Tensor gradWeight, // at::Tensor gradBias, + at::Tensor columns, at::Tensor ones, int kW, int kH, int dW, int dH, + int padW, int padH, int dilationW, int dilationH, int group, + int deformable_group, float scale, int im2col_step); + +void modulated_deform_conv_cuda_forward( + at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones, + at::Tensor offset, at::Tensor mask, at::Tensor output, at::Tensor columns, + int kernel_h, int kernel_w, const int stride_h, const int stride_w, + const int pad_h, const int pad_w, const int dilation_h, + const int dilation_w, const int group, const int deformable_group, + const bool with_bias); + +void modulated_deform_conv_cuda_backward( + at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones, + at::Tensor offset, at::Tensor mask, at::Tensor columns, + at::Tensor grad_input, at::Tensor grad_weight, at::Tensor grad_bias, + at::Tensor grad_offset, at::Tensor grad_mask, at::Tensor grad_output, + int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h, + int pad_w, int dilation_h, int dilation_w, int group, int deformable_group, + const bool with_bias); +#endif + +int deform_conv_forward(at::Tensor input, at::Tensor weight, + at::Tensor offset, at::Tensor output, + at::Tensor columns, at::Tensor ones, int kW, + int kH, int dW, int dH, int padW, int padH, + int dilationW, int dilationH, int group, + int deformable_group, int im2col_step) { + if (input.device().is_cuda()) { +#ifdef WITH_CUDA + return deform_conv_forward_cuda(input, weight, offset, output, columns, + ones, kW, kH, dW, dH, padW, padH, dilationW, dilationH, group, + deformable_group, im2col_step); +#else + AT_ERROR("deform conv is not compiled with GPU support"); +#endif + } + AT_ERROR("deform conv is not implemented on CPU"); +} + +int deform_conv_backward_input(at::Tensor input, at::Tensor offset, + at::Tensor gradOutput, at::Tensor gradInput, + at::Tensor gradOffset, at::Tensor weight, + at::Tensor columns, int kW, int kH, int dW, + int dH, int padW, int padH, int dilationW, + int dilationH, int group, + int deformable_group, int im2col_step) { + if (input.device().is_cuda()) { +#ifdef WITH_CUDA + return deform_conv_backward_input_cuda(input, offset, gradOutput, + gradInput, gradOffset, weight, columns, kW, kH, dW, dH, padW, padH, + dilationW, dilationH, group, deformable_group, im2col_step); +#else + AT_ERROR("deform conv is not compiled with GPU support"); +#endif + } + AT_ERROR("deform conv is not implemented on CPU"); +} + +int deform_conv_backward_parameters( + at::Tensor input, at::Tensor offset, at::Tensor gradOutput, + at::Tensor gradWeight, // at::Tensor gradBias, + at::Tensor columns, at::Tensor ones, int kW, int kH, int dW, int dH, + int padW, int padH, int dilationW, int dilationH, int group, + int deformable_group, float scale, int im2col_step) { + if (input.device().is_cuda()) { +#ifdef WITH_CUDA + return deform_conv_backward_parameters_cuda(input, offset, gradOutput, + gradWeight, columns, ones, kW, kH, dW, dH, padW, padH, dilationW, + dilationH, group, deformable_group, scale, im2col_step); +#else + AT_ERROR("deform conv is not compiled with GPU support"); +#endif + } + AT_ERROR("deform conv is not implemented on CPU"); +} + +void modulated_deform_conv_forward( + at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones, + at::Tensor offset, at::Tensor mask, at::Tensor output, at::Tensor columns, + int kernel_h, int kernel_w, const int stride_h, const int stride_w, + const int pad_h, const int pad_w, const int dilation_h, + const int dilation_w, const int group, const int deformable_group, + const bool with_bias) { + if (input.device().is_cuda()) { +#ifdef WITH_CUDA + return modulated_deform_conv_cuda_forward(input, weight, bias, ones, + offset, mask, output, columns, kernel_h, kernel_w, stride_h, + stride_w, pad_h, pad_w, dilation_h, dilation_w, group, + deformable_group, with_bias); +#else + AT_ERROR("modulated deform conv is not compiled with GPU support"); +#endif + } + AT_ERROR("modulated deform conv is not implemented on CPU"); +} + +void modulated_deform_conv_backward( + at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones, + at::Tensor offset, at::Tensor mask, at::Tensor columns, + at::Tensor grad_input, at::Tensor grad_weight, at::Tensor grad_bias, + at::Tensor grad_offset, at::Tensor grad_mask, at::Tensor grad_output, + int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h, + int pad_w, int dilation_h, int dilation_w, int group, int deformable_group, + const bool with_bias) { + if (input.device().is_cuda()) { +#ifdef WITH_CUDA + return modulated_deform_conv_cuda_backward(input, weight, bias, ones, + offset, mask, columns, grad_input, grad_weight, grad_bias, grad_offset, + grad_mask, grad_output, kernel_h, kernel_w, stride_h, stride_w, + pad_h, pad_w, dilation_h, dilation_w, group, deformable_group, + with_bias); +#else + AT_ERROR("modulated deform conv is not compiled with GPU support"); +#endif + } + AT_ERROR("modulated deform conv is not implemented on CPU"); +} + + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("deform_conv_forward", &deform_conv_forward, + "deform forward"); + m.def("deform_conv_backward_input", &deform_conv_backward_input, + "deform_conv_backward_input"); + m.def("deform_conv_backward_parameters", + &deform_conv_backward_parameters, + "deform_conv_backward_parameters"); + m.def("modulated_deform_conv_forward", + &modulated_deform_conv_forward, + "modulated deform conv forward"); + m.def("modulated_deform_conv_backward", + &modulated_deform_conv_backward, + "modulated deform conv backward"); +} -- cgit v1.2.3